Radiomics Approaches in Gastric Cancer: A Frontier in Clinical Decision Making
Gastric cancer remains a significant global health burden, despite a decline in its incidence over recent decades. It is the third leading cause of cancer-related deaths worldwide, with over 1,000,000 new cases and approximately 783,000 deaths reported in 2018. In China alone, the numbers were staggering, with 6,791,000 new cases and 498,000 deaths in 2015. Imaging modalities play a crucial role in the diagnosis, staging, and risk stratification of gastric cancer, aiding in the selection of optimal therapeutic strategies and improving patient outcomes. Radiomics, an emerging field, has shown promise in extracting quantitative features from medical images that are not visible to the naked eye or quantifiable through routine analysis. This article reviews the application of radiomics in gastric cancer, its challenges, and future prospects.
Overview of Radiomics
Radiomics was first introduced by Lambin et al. in 2012 and later refined by Kumar et al. as the high-throughput extraction and analysis of large amounts of advanced quantitative imaging features from medical images obtained via computed tomography (CT), positron emission tomography (PET), or magnetic resonance imaging (MRI). The primary advantage of radiomics is its ability to acquire numerous quantitative features that provide information on tumor phenotype and microenvironment, which are unavailable through traditional radiology. Another significant strength of radiomics is its use of artificial intelligence and machine learning techniques to transform high-dimensional data into diagnostic, predictive, or prognostic models or signatures that support personalized clinical decision-making.
A typical radiomics study is structured into four phases:
- Image Acquisition: Obtaining large-scale medical images with standard scanning and reconstruction protocols is essential to eliminate unnecessary variability.
- Image Segmentation: Regions of interest (ROIs) or volumes of interest (VOIs) of the tumor, metastatic lesions, and normal tissues are segmented manually or semi-automatically for further analysis.
- Feature Extraction and Selection: High-throughput extraction of quantitative imaging features from ROIs or VOIs is the essence of radiomics. Commonly used features include shape and size features, first-order histograms, second-order histograms (textural), and fractal features. Redundant or irrelevant features are excluded using methods like LASSO, maximum relevance and minimum redundancy, and principal component analysis.
- Model Construction and Validation: Optimal machine-learning models are identified based on clinical information and selected features. Techniques such as support vector machine (SVM), random forest, artificial neural networks (ANNs), and bootstrapping are widely used. Models must be validated before application in scientific and clinical communities to ensure statistical consistency between training and validation sets.
Application of Radiomics Approaches in Gastric Cancer
A comprehensive search of PubMed databases using terms like “radiomics,” “texture analysis,” and “gastric cancer” yielded 17 original articles. These studies demonstrated that radiomics has moderate to excellent performance in various aspects of gastric cancer management, including differential diagnosis, histological grade assessment, tumor staging, prediction of therapy response, and prognosis.
Differential Diagnosis
Distinguishing between primary gastric lymphoma, gastrointestinal stromal tumor (GIST), and adenocarcinoma is challenging due to their similar appearances on routine CT. Two studies explored the use of radiomics for differential diagnosis. Ba-Ssalamah et al. analyzed texture features from pre-operative arterial and portal phase CT images, finding that VOI-based texture features from arterial phase images could differentiate GIST from lymphoma with 100% accuracy and adenocarcinoma from lymphoma with a misclassification rate of 3.1%. Ma et al. reported that whole-lesion-based texture features from portal phase CT images could differentiate adenocarcinoma from lymphoma with an accuracy of 87%.
Prediction of Histological Grade
Histopathological features significantly influence treatment and prognosis. Liu et al. segmented whole lesions on pre-operative arterial and portal phase images of 107 patients, identifying radiomic features correlated with histological grade and Lauren type. Zhang et al. collected ADC maps of 78 patients and found that histogram parameters differed significantly between lesions with different histological grades, though their clinical utility was limited due to low AUC values.
Prediction of Tumor Stage
Accurate tumor staging is essential for selecting appropriate therapeutic approaches. Five studies evaluated radiomics for predicting lymph node status, vascular invasion, and occult peritoneal metastasis. Liu et al. found that whole-lesion-based radiomic features could identify positive lymph node metastases with 74-81% accuracy but were not useful for predicting T stage. Feng et al. developed a radiomics model using pre-operative portal phase CT images of 490 patients, achieving an AUC of 0.824 in the training cohort and 0.764 in the test cohort for predicting lymph node metastasis. Dong et al. conducted a multi-center study on 554 patients with occult peritoneal metastasis, developing a nomogram with an AUC of 0.958 in the training set and 0.941, 0.928, and 0.920 in validation sets.
Prediction of Response to Therapy and Patient Prognosis
Identifying pre-therapeutic predictive markers for response and prognosis is invaluable for individualized treatment.
Prognosis of Surgical Resection
Giganti et al. investigated the association between CT texture-derived parameters and overall survival (OS) in 56 patients, finding that features like energy, entropy, and skewness were significantly associated with poor prognosis. Li et al. developed a nomogram incorporating ROI-based radiomics signatures and clinical parameters, which provided better predictive accuracy for prognosis than either radiomics signatures or clinical parameters alone.
Prognosis of Neoadjuvant Chemotherapy (NAC)
Four studies explored radiomics for predicting response and prognosis of NAC. Giganti et al. found that entropy, range, and root mean square were independent predictors of response. Jiang et al. conducted a multi-center study on 1591 patients, developing a prognostic classifier with hazard ratios (HRs) of 2.98, 3.17, and 2.671 for predicting disease-free survival (DFS) and 3.72, 3.415, and 2.830 for predicting OS in training, internal, and external test sets, respectively.
Prognosis of Targeted Chemotherapy with Trastuzumab
Yoon et al. studied 26 patients with HER2 overexpression, finding that GLCM features like contrast, variance, and correlation could differentiate responders from non-responders with AUCs ranging from 0.75 to 0.77.
Prognosis of Radiotherapy
Hou et al. segmented VOIs of pre-treatment arterial phase CT images of 43 patients, finding that radiomics signatures could predict response to radiotherapy with AUCs of 0.714 and 0.749 using ANN and k-nearest neighbor methods in the training set, and 0.816 in the validation set.
Challenges and Future Prospects
Despite its promise, radiomics in gastric cancer is still in its infancy, facing several challenges.
Image Acquisition
The power of radiomic models depends on sufficient patient populations. Most studies are retrospective, with images from multiple scanners and variable slice thicknesses, introducing variability in features. Standardized imaging protocols are recommended to improve reproducibility.
Image Segmentation
Variability in segmentation methods can introduce bias. Computer-aided edge detection followed by manual curation is currently the best method, though deep learning approaches may offer improvements. The efficiency of 3D VOI-based features is compromised by partial volume artifacts, and their superiority over 2D ROI-based features requires further study.
Feature Extraction and Selection
Various software and programs are used for feature extraction, and not all features are useful. The process of feature reduction should be documented clearly to ensure robustness.
Model Construction and Validation
Different modeling techniques have inherent limitations, and the choice of technique affects prediction performance. Models should be externally validated to ensure generalizability.
Future Direction of Radiomics in Gastric Cancer
Future research should focus on the long-term survival prediction after NAC, identification of specific cancer sub-groups for targeted therapy, and the use of multimodality imaging-based radiomics for a more complete picture of the tumor.
Conclusions
Radiomics represents a frontier in clinical decision-making for gastric cancer, offering advanced algorithms to extract valuable information from medical images. While still in its early stages, radiomics has the potential to revolutionize precision medicine in gastric cancer with continued data accumulation, workflow standardization, and advancements in artificial intelligence techniques.
doi.org/10.1097/CM9.0000000000000360
Was this helpful?
0 / 0